KVM Forum 2021

The current qemu auto-converge implementation works by throttling all the vcpus based on dirty-rate and network-throughput observed over an iteration. This leads to - all the vcpus getting equally penalized, multiple iterations to get to the optimal throttle value, and convergence failure when >99% throttle is required. This talk proposes a new algorithm based on individual vcpu “dirty quota” that dynamically throttles individual vcpus based on their contribution to overall memory dirtying. It does so by enforcing an upper limit(dirty-quota) on amount of pages a vcpu is allowed to dirty within a small interval of time to limit the total pages dirtied over an iteration. The advantage of this approach is that it penalizes only the write-intensive vcpus, dynamically adapts in case of varying network and dirty-rate and improves convergence by allowing more granular control over dirty rates.